[00:00:00] David Bergvinson: Well, welcome to the third episode of Grounded Intelligence. And in this podcast, we’re looking at the responsible use of AI. Now, you recall that in episode two, we talked about data, which is really the fuel for artificial intelligence. In this episode, we’re gonna be talking about models and how to make sure that they’re relevant for the needs of farmers and other actors along the agriculture value chain.
[00:00:23] David Bergvinson: So it’s gonna be a very important and exciting topic, and I’m looking forward to diving into this one. And so what we’re gonna be looking at around models is that they are, in words, biased. Depends on what you feed them in their development that impacts on how they deliver, uh, recommendations or outputs.
[00:00:43] David Bergvinson: This becomes very important when we consider the fact that a lot of the, what we call frontier models, have been developed on data from, uh, North America, Europe, and largely in English. And so now what we need to be thinking about is, well, does that create a bias towards delivering [00:01:00] relevant solutions to farmers in places like Africa or Asia?
[00:01:03] David Bergvinson: And so this is really the tension we have right now in this rapidly evolving world of AI is can we use those frontier models for the applications related to agriculture in Africa or Asia, or do we need specifically built, what we sometimes call as small or specific models, for Africa and Asia to ensure that the recommendations are relevant for farmers in those geographies?
[00:01:29] David Bergvinson: And so with me to talk about this really important topic are three colleagues from Dev Afrique, uh, to help guide us, uh, through this conversation and unpack some of the issues that were, uh, discovered during the development of a discussion paper on models. And so joining me is Femi Royal. He’s a senior advisor helping with governance and systems perspectives, uh, for development.
[00:01:54] David Bergvinson: Malcolm Durosaye, who’s a co-author of the AgX AI discussion paper. He’s an, [00:02:00] uh, consultant with Dev Afrique. And Ahmad Raji, who is also a co-author of the AgX AI discussion paper and an associate, uh, consultant with Dev Afrique. And so welcome, gentlemen. Um, I would just invite you to amplify a little bit the introduction, uh, especially as it relates to the role models play in your own sphere of work as a consultant in Africa
[00:02:24] David Bergvinson: And so just to amplify on those introductions, I invite Femi Royale to give us a little bit more perspective of, uh, his domain and the relevance of models in supporting his work at DevAfrik. So Femi
[00:02:41] Femi Royal: Thank you very much, uh, David. Um, like you mentioned, I’ve been– I’m consulting with, uh, DevAfric. Uh, to a large extent, I’ve been working on, you know, digitizing most of the agriculture engagement within the Nigerian space, um, largely [00:03:00] because, you know, we’ve got very fragmented, uh, farmers’ data, and we had to collect quite a lot of data across the country so we could, um, get a bit of more insight into what’s been happening and how we can make some improvement within the agricultural value chains to tailor, you know, um, some of the solutions that have been created to the farmers’ needs overall.
[00:03:24] Femi Royal: And I think that that’s been very interesting and rewarding so far, and, uh, I’m looking forward to share more about it as we, as we navigate this conversation. Thank you. Yeah. Great. And, uh, the contextual relevance of these models is so important for that kind of work. Um, Malcolm, I invite you to give us a little bit more background on your work and how you see the application of these models specifically for, uh, your line of work in Africa.
[00:03:52] Malcolm Durosaye: Thanks, David. Um, my name is Malcolm Dush. I am a consultant with DevAfric Development Advisors, and I [00:04:00] lead the, uh, discussion paper on localized Agri LLM exploring low power, low costs models for SSPs, which is the small-scale producers in Africa and India. And my work fo- focuses on the, the section of AI, data transformation, and data systems, and particularly on the emerging technologies which includes AI.
[00:04:28] Malcolm Durosaye: And through this research, we’ve explored the opportunities and challenges of localizing AI for agricultural advisory services with a focus on issues such as DPIs, farmer data governance, inclusion, sustainability, trust, and accountability. I’m pleased to join, uh,
[00:04:51] Malcolm Durosaye: this conversation. Great. This is becoming increasingly important as countries across Africa and Asia, you know, explore how [00:05:00] AI can create meaningful value for smallholder farmers at scale. Thank you. Thank you. Yeah, you mentioned the word trust, and we’re gonna circle back to that word later in our conversation, but very important.
[00:05:11] David Bergvinson: Ahmad, would you, uh, please, uh, give us a little bit more context about yourself and the role that models play in your work? All right. Yeah. Thanks, David. My name is, um, Ahmad Raji, and I’m associate consultant with, um, DevAfric. So, um, I’m also a co-author for the, um, discussion paper, the, um, localized Agri LLM
[00:05:33] Ahmad Raji: for small-scale agricultural producers in, um, Africa and in, um, in India. So, um, some of the interesting aspect that I’ve seen our models work in the context of agriculture and some of the things we’ve highlighted in our discussion paper is, um, models have the ability to, um, integrate information from different source- sources, um, aggregate them, and then present them into a meaningful and, um, a path that, um, end users can easily follow to achieve their required goals.
[00:05:58] Ahmad Raji: So in the context of, in the [00:06:00] context of agriculture, um, users that are not very vast in some particular topics, they can easily distill information that is necessary for them. And then with that information, they can have, um, particularly, um, different, um, outputs in terms of probably increasing yield or even, um, combating infestation, pest infestation, and some other challenges like that.
[00:06:20] Ahmad Raji: So models is very important. Models, they can bring all this information at the fingertips of the farmer, so they don’t need to go around to check from other sources or even have a physical interaction with extension agents, particularly
[00:06:34] David Bergvinson: Great. Well, thank you all. Um, and so, uh, just diving into our conversation, we talked a little bit in the introduction around, uh, the context or relevance of these large language models, uh, and especially when we talk about existing frontier models. But they fall short in understanding the context quite often because we don’t have the data from a specific country, let’s say Nigeria, uh, to really populate [00:07:00] them and fine-tune them for, uh, application in a specific country like Nigeria.
[00:07:05] David Bergvinson: So Femi, uh, given your role in policy, uh, you know, what does li- uh, you know, localization mean for you as we talk about the contextualization of these models?
[00:07:17] Femi Royal: Uh, thank you David for that question. Um, for me, I think there’s, um, there’s a lot to unpack there. Um, first of all, we have to be thinking of things from where we’re getting, where we’re getting the conversation wrong. Um, particularly from where we sort of see some form of misconception. I think the biggest misconception that I have seen at, uh, within the AI space is ge- generally the assumption that, uh, localization problem within the agricultural context in Africa, maybe in Asia, it’s, it’s more a language issue.
[00:07:54] Femi Royal: Uh, but, but to a large extent, if we, if we dig into that a little bit more, we realize that language issue [00:08:00] is pretty much, you know Looking at the issue from a surface level, uh, and that doesn’t get us into what the root cause really is. Um, ’cause if you look at agriculture, just like you mentioned, there’s, uh, a lot of context.
[00:08:15] Femi Royal: Uh, it’s a context-dependent sector, um, all over the world, not just in, in Nigeria, uh, or in Africa. It’s all over the world. In the US, in Europe as well, it’s very context-specific. And what applies to a farmer in Europe, what applies to a farmer even in India, for example, because we’re looking at the Asian context, um, it’s probably very irrelevant to, um, a rice farmer or a cassava farmer in Nigeria, right?
[00:08:43] Femi Royal: Because the context is very different. So I think the conversation is pretty much supposed to be targeting, um, more, uh, in adapting to the intelligence o- of the local realities that exist. So, and that’s where, you know, paying more attention to [00:09:00] context really matters. Uh, and this, this context actually generates quite a lot of data, literally.
[00:09:05] Femi Royal: And models that are going to be dev- uh, deployed need to be getting data, n- need to be feeding the, the data around, um, the context, especially, for example, in Nigeria. The context around local crops, around local weather patterns, around local soil conditions, around local pest outbreaks, around local market dynamics as well, or even local farmer, farming practices.
[00:09:31] Femi Royal: So I feel like if models don’t have enough data around this context, um, it’ll be difficult for them to, uh, especially when you’re training them, it’ll be diff- difficult for them to give recommendations that are intelligent and apply or even doesn’t sound disconnected to the realities of the farmer operating in that context.
[00:09:53] Femi Royal: So like, like I said earlier, um, we, we mustn’t, uh, fall into the trap of assuming that [00:10:00] language is what localization’s about. W- we need to be thinking more from, um, training, uh, models around the context of the environment where we’re speaking about. And like I said earlier, the– there’s, it’s a, there’s a lot of culture, there’s a lot of language, there are lots of knowledge systems, and the data comes out from all of these bits and bobs, and they need to be getting into the models so that the models can actually Uh, provide, you know, uh, m- be able to address the needs of the farmer directly.
[00:10:35] Femi Royal: Uh, and that’s how we get to build trust. I’m sure you’re probably gonna be touching on trust later in the conversation, but, um, uh, and, and that’s how, uh, that’s what I’m thinking, right? Because… And, and just to close this, this part of the conversation, I, I think that if you look at the context of India, for example, um, they’ve been doing a lot of work building their digital infrastructure from a foundational level using, um, like, uh, tools [00:11:00] like Bashny, Agric Stack, you know, and they’re sort of making, uh, creating that platform where there’s some form of localized innovation.
[00:11:09] Femi Royal: But, you know, if you think about Africa, I think there’s still that gap where innovation hasn’t really gotten its foundation built enough, especially around models really getting, uh, fed with enough data around the context where they’re operating. Um, and, and that’s one thing we probably need to do more work on as we, as we get onto, you know, on this particular journey.
[00:11:33] Femi Royal: Uh- Agreed … I’ll close it off for now. Thank you. Thank you. Well, well, let’s, uh, take off from there, uh, Malcolm, to talk a little bit about this comparison between Africa and India. You know, India-
[00:11:47] David Bergvinson: Digital, digital public infrastructure. And I, I want to get your perspective on how that translates to the African context Thanks. [00:12:00] Um, I think one thing that became very clear when we were doing this research is that African innovators are not waiting for the perfect infrastructure before deploying AI.
[00:12:12] Malcolm Durosaye: Um, they are still trying to navigate different challenges, trying to build their own database, um, systems and all that, um, across. But we’ve seen organizations use voice-based, um, systems where literacy is a barrier in one instance, and then we’ve seen organizations use SMS business D where, uh, smartphone ownership is limited.
[00:12:35] Malcolm Durosaye: And we’ve also seen organizations use extension agents as trust intermediaries between technology and farmers. And, um, one of the example is, uh, the PlantVillage Noroo. Uh, we also have FarmChat, FarmerChat. We have, uh, Kilimo by ITA and others that designed this, [00:13:00] uh, operation in low connectivity environment and just to reach the farmers at the last mile.
[00:13:06] Malcolm Durosaye: Um, this, this approach are really important, uh, as they demonstrate that innovation can happen even in low resource, um, environment. But they also highlight an important reality. There’s a limit to how far they can actually work around these, um, challenges. Uh, while these solutions can, um, you know, navigate some of these challenges that we have, um, currently on the continent, they don’t un– address the underlying challenges of fragmentation.
[00:13:41] Malcolm Durosaye: We still have issues around data silo across different ministries, across different, um, organization. And too often, organizations are collecting the same data repeatedly across different, uh, use cases, maintaining se-separate farmer databases, and operating in [00:14:00] system that cannot easily, um, exchange information or services.
[00:14:05] Malcolm Durosaye: They are not interoperable. And then this is where the contrast between India and Africa, you know, comes in. Uh, we’ve seen the likes of AgriStack, we’ve seen the likes of Bashiny, uh, which addresses issues around language barriers across, um, you know, farmer registries and all that. And in Africa, most of the African innovators are currently navigating and also building their own, you know, individual APIs to address some of
[00:14:38] Malcolm Durosaye: these challenges. So that’s what we have, um, you know, across both contexts that we’ve, you know, assessed through this paper. Yeah, great. And we’ll dive into more details on that important topic because, um, you know, India now is developing its own sovereign large language models. I, I believe there’s [00:15:00] seven of them that are being developed concurrently, uh, with a large emphasis on the agriculture sector, given that agriculture constitutes such a large portion of the population, much like it does in Africa.
[00:15:10] David Bergvinson: So it’ll be interesting to see them on that journey and what could be adapted to the African context or, or maybe even, um, adopted. Um, so, uh, let me pivot a little bit to this whole issue of a- adoption. Um, you know, there are tools out there, and so maybe, Ahmad, you could speak a little bit more, uh, to the tools that you researched, uh, in the discussion paper and what sort of came out to you as around barriers to adoption and what could p- potentially limit the impact of AI in supporting farmers in Africa.
[00:15:45] Ahmad Raji: Yeah. Thank you very much. So, um, when, when, when you create, um, AI models and you expect it to be delivered to the last mile, to the small-scale producers, you need a solid plan on how to deliver those models to the, um, to the last mile, to people in the [00:16:00] rural areas for them to be able to adopt the model.
[00:16:02] Ahmad Raji: And for you to be able to achieve that, you, um, you need a structure, a structure in place. So you need like, um, a digital structure in place on which the AI innovation will be built upon. And, um, in, in, in my own opinion, I think when you’re, when you’re, when you’re assessing the success of an AI, you don’t just talk about the success at, um, the startup stage when, um, when you’re piloting AI.
[00:16:25] Ahmad Raji: You need to also think about it in terms of sustainability. Is the infrastructure able to support the innovations a- at, at this point? Take for example, um, Akilimo in Nigeria and Tanzania. Akilimo is one of, um, the most adopted, um, AI platforms in, um, in Nigeria. And what they’ve been able to achieve is, um, like they were able to, um, to bring a farmer ecosystem together.
[00:16:45] Ahmad Raji: So they didn’t exclude the farmer ecosystem, the structure of the farmers. And then they are delivering-[00:17:00] [00:18:00]
[00:18:10] David Bergvinson: Yeah. I, I think you raise a good point there around trust and the role that, I would call them lead farmers in a community play, in socializing new technology and demonstrating its value to their community. So yeah, there’s an anthropology side to adoption that we really need to take on board as we think about r- you know, building and retaining trust of farmers, uh, the context of these models so they deliver value to farmers and mitigate risk.
[00:18:35] David Bergvinson: So there’s a lot here and, and so as we develop these models for the African agriculture context, we need to really keep all of this in mind, so thank you for that. Um, so just building on this topic of trust, which is a word that’s come through in each one of our episodes so far, um, let’s dive into that a little bit more because, you know, as we work in these, in the AI era, there is a lot of concern [00:19:00] around misinformation.
[00:19:02] David Bergvinson: Uh, early models had issues of hallucination, for example, and, you know, generating, you know, I won’t say nonsense, but non-referenced, uh, insights that, you know, people were accepting as truth, but in reality weren’t. And so building that trust and, and your point, Ahmad, about, uh, extension workers, lead farmers as a bridge to that, that building that trust.
[00:19:23] David Bergvinson: I’d just like to explore this a little bit further. So like, Femi, as, as we talk about, uh, trust in the context of, um, you know, not betraying, uh, the trust of farmers, like not giving false advice, uh, how, how do you see this unfolding from a policy perspective when it comes to issues of liability, where a message may go out that’s not accurate, a farmer takes action based on that, and, and there’s a negative consequence?
[00:19:50] David Bergvinson: Are governments thinking through this liability from a policy perspective currently? Um, thank you for that question. [00:20:00] Uh, so first of all, I, I like to say that, yeah, we, we, we’ve seen the way that agri LLMs have been evolving over the years, uh, thanks to efforts of, from organizations like OpenAI and Anthropic and other, you know, DeepSeek AI and the likes who are making a lot of advancement within the space, creating very powerful models, um, you know, and helping us have like multi-modal reasoning capabilities all along.
[00:20:33] Femi Royal: But if, if we, if we start to think about it, we’ll realize that, yeah, just as you highlighted, um, AI has become buzzword of the season, and LLM seem to be like a very important component of it, and it has ultimately made that even become a strategic focus for a lot of government organizations. I mean, not to talk of industries and research institutions.
[00:20:59] Femi Royal: [00:21:00] Uh, it– I know in the paper we did mention around the fact that there are more than 70 countries who have, um, created, uh, sort of like AI strategy, uh, national AI strategy actually, that seem to cover concept or themes around data privacy, ethics, um, innovation and the likes, right? Uh, but again, th-this is, uh, however, bearing in mind that, um, every innovation has also got its, um, pro– its, its drawbacks or its challenges, and one of which is the hallucination and, and the, I mean, the disclaimer that we get to see that, yeah, every informatio-information, I mean, when we use our regular LLMs like the ChatGPTs, we definitely get, um, told that, yeah, they could create, they could produce erroneous information.
[00:21:49] Femi Royal: And that in itself is a, is a huge risk, you know, uh, in terms of the, the quality of information that gets out to farmers. Uh, but I [00:22:00] think You know, the big, the big, um, the big AI companies are majorly to the, to a large extent, I think more, more of them are focusing on high compute, high resource, you know.
[00:22:13] Femi Royal: But the divide for me is actually where I’m even gonna be focusing more on, which is still tying to access and inclusion for, um, smaller farmers in LMICs, I mean the low and medium income, um, countries, especially in Africa, which is our, one of our focus country, uh, continents, right? Uh, and, and I think if we go even a little bit down into the, into the conversation, we’ll, we’ll see that there’s a bit of bias as well that I, that is created in the module.
[00:22:44] Femi Royal: And I tend to think that to a large extent the bias is more from an accessibility point of view because a model can be trained on different types of data, but it still excludes the farmers, you know. [00:23:00] And at, at the end of the day, we tend to assume that farmers have got reliable internet, they’ve got smartphones, they’ve got fluency and all that stuff, right?
[00:23:12] Femi Royal: But the opposite is the reality, right? You know, before we even start talking about the output of the AI itself, I feel like there’s a huge gap in the fact that farmers don’t currently, especially farmers in LMICs, don’t currently even have enough accessibility to be able to tap into the resources or the capabilities of agri AI.
[00:23:36] Femi Royal: So while the model works and I, and I think that, you know, government is trying to do a lot, like I said earlier, around data protection, around, um, ensuring that some of these things are well trained, uh, and they really, there’s some form of liability as well, um, across board. But I, I feel like, you know, there’s still, um, accessibility inequality [00:24:00] that is plaguing, um, local agricultural farmers.
[00:24:04] Femi Royal: And that is because when at the point of design, as-assumptions have, uh, sort of included re- a, a different reality of the current experience of local farmers. So for me, I think it’s more, the conversation is really more about, you know- driving inclusion for farmers, especially from the start, you know, across language, across delivery channels, um, and making sure that, you know, uh, we don’t have a module failure as a result of poor design, you know, a-and that we have a bit more technically accurate outputs from, um, the LLMs that we eventually adopt with our local farmers.
[00:24:47] David Bergvinson: Great points. Yeah. I, I, I sometimes refer to the AI era as, uh, digital divide 2.0. You know, with the internet we had the first digital divide, and now with AI, we run the risk of a second and pro- potentially larger [00:25:00] digital divide, uh, with emerging markets. And so, uh, the points you raise are very relevant. Um, you know, one of the things that, uh, always is in the back of my mind is when we make investments through, uh, you know, donor organizations to develop these resources, uh, in this case, you know, advisory services, for example, in Africa, we’re not always thinking about sustainability in the business model to support these ongoing.
[00:25:28] David Bergvinson: Amit, uh, as you went through the, the model discussion paper, what was going in your mind around sustainability and, and how can we, um, design these with sustainability in mind?
[00:25:53] David Bergvinson: Yes, and he’s, uh
[00:25:58] David Bergvinson: Okay, let me just pivot just to keep [00:26:00] momentum going with Malcolm then. Um, yeah, so, so Malcolm, those are really great points and, and one other issue that comes to my mind is the issue of sustainability. And you’ve talked a lot about the context. What is, in your mind, the path towards sustainable development of these resources after, you know, initial investor or donor money runs dry?
[00:26:22] David Bergvinson: How do we keep these services, uh, viable in the long term?
[00:26:30] Malcolm Durosaye: So you can come, come over here
[00:26:37] David Bergvinson: Yeah
[00:26:41] David Bergvinson: Uh, Malcolm, are you there? Yeah, I can hear you. Okay. So I’m, I’m redirecting the question to you, uh, ’cause we’re having problems with Ahmad’s mic. Um, so, so I’m just gonna start again. So Ahmad, uh, these are great points that you raise around context. You know, another big problem or challenge we see is [00:27:00] around sustainability of these types of services, AI-enabled services, especially, uh, advisory services.
[00:27:06] David Bergvinson: What in your mind do you see as a sustainable model for these resources, uh, going forward in Africa? Uh, so, uh, I think one of the reoccurring pattern we’ve seen, um, in agricultural technology is that innovation and sustainability are not always the same thing most times. And we’ve seen, um, with interviews with different stakeholders where many promising tools are launched with, you know, grant funding.
[00:27:39] Malcolm Durosaye: They de-demonstrate strong results during pilot phases, and farmers use them, such as also celebrates them, you know, um, choose them across board, test them. But when the funding ends, uh, the question now becomes: Who pays for this solution going forward? And one of [00:28:00] the challenges, um, is that most smallholder farmers operates with very thin margins, um, even if they too creates value, um, expecting farmers to bear the full cost, you know, at that length, uh, which most times it’s very difficult because, uh, they have to work around what the farmer can actually afford.
[00:28:25] Malcolm Durosaye: So at the same time, you know, they’re also relying on donor funding in Africa– particularly in Africa, which, um, they feel like should sustain them for a long time, which, you know, in reality, we realize that it’s not really sustainable. We could see the, um, an example with, uh, you said, right? And at the same time, we could also see from different funding where, you know, it only function at the pilot phase.
[00:28:56] Malcolm Durosaye: And that’s why we think that the [00:29:00] future of Ag AI, uh, lies in blended model, where government can provide foundational, um, public, you know, infrastructure like DPI, like AgriStart, Pashimi, and the likes in India. And then private sector, um, can also build services on top of those, uh, foundations and create commercial value to the farmers.
[00:29:24] Malcolm Durosaye: Yeah. And then the development partner can also support with innovation, you know, inclusion, and early stage, um, risk-taking for the solution providers. Well, I can- We’ve seen that, um, across board, and we, um- You know, spoke with a lot of stakeholders within the sector just to find out the best model that could enable sustainability in a very long term.
[00:29:50] Malcolm Durosaye: And we realized that it’s, uh, one B2G, uh, solution providers, um, providing services to government and also B2B, which is like, uh, solution providers, um, providing so- um, services to the businesses across board. So those are[00:30:00]
[00:30:12] Malcolm Durosaye: the most sustainable, uh, business model that could work, you know, in very long term. Well, in Africa, the, the telcos actually are likely to play a very large role. So, uh, you know, if I look at, you know, the mobile money era with M-PESA and Safaricom in Kenya, you know, it was the telcos that actually did a lot of that, you know, investment lift to, uh, launch these services.
[00:30:39] David Bergvinson: And I think the same will be true even in the AI era, where telcos will play a significant role, especially in, in, uh, African countries where government budgets are str- constrained and, um, you know, telcos already have a relationship with that end user, the farmer. So it’ll be interesting to see how that unfolds.
[00:30:58] David Bergvinson: Um, I’d [00:31:00] just like to go now to, uh, Ahmad just to close us out on some of the issues that are going to be facing us as we look at data and its generation, ownership, and application for LLMs. Um, from your experience, you know, what do you see from the farmer’s perspective around data and, and ownership? And then I- I’ll, I’ll pivot, um, to Femi to give us the government perspective.
[00:31:27] David Bergvinson: So, uh, first, first to you, uh, Ahmad. All right. Thank you very much. And, um, this is one of the defining questions for the future of, um, digital agriculture. And for a very long time now, our conversation about, um, farming has usually been centered on, uh, on the land, the inputs, the market, and then we have very few discussions around, uh, around data because, um, data has, um, has become an asset in its own right, and many people have shied away from, from that discussions.
[00:31:59] Ahmad Raji: But, um, consider [00:32:00] what happens when a farmer uses a digital platform. They, um, typically upload pictures of themselves, pictures of their farm, of their assets, even machineries to inform, um, what kind of farming practices would support those machineries and, and things like that. So AI can train on this data, create models on them, a-and the likes.
[00:32:19] Ahmad Raji: But, um, there are some critical gaps that needs to be, to be addressed. And much of the problem is that, um, most farmers have no visibility into what happens next, what happens to their, to their data, how their data is being used behind the scenes. They, um, some of them only gets the, um, the outputs from, from the models.
[00:32:35] Ahmad Raji: But what happens with the data, they rarely know where their data is being stored, who has access to it, and they have no control over it. So, um, having control over, over, um, farmer data is, um, is kind of a way of correcting that imbalance. And in practice, it rests on some few principles that needs to be, to be addressed, and the first one is, um, transparency.
[00:32:55] Ahmad Raji: So everybody within the value chain, within the pipeline needs to know where they belong. [00:33:00] They need to know the data that they have been, um, they’re being generating for, um, the ecosystem. They need to know where it’s going, who has the data at every point, what– to what extent are they able to use the data.
[00:33:10] Ahmad Raji: And that in a way also feed back into trust because when you know your data won’t be used against you, you kind of trust the system, and then you also want to, want to use it limitless. You want to give as much, um, context that, um, that you have, that you can provide for, for, for the model. And then the second is also consent.
[00:33:26] Ahmad Raji: So they need to, um, explicitly state some of the things that they want their data to be used for. So, um, they should have meaningful choices. And some-sometimes some of these, um, small-scale producers, they might not really understand the complexity behind some of, um, these phrases in terms and conditions and, and things like that.
[00:33:43] Ahmad Raji: So you would also need a kind of, um, a person that would explain to them in languages that they understand, in realities that are, um, in, in context that reflects their realities, so that they will understand what they are consenting to. And then in future, when they see, um, when they see things that are made, been, been made from the data, they understand that this [00:34:00] is something that they have envisioned right from the beginning, and then it, it, it builds the trust and, and the likes.
[00:34:05] Ahmad Raji: And then the third is, um, for, um, value sharing also. So sometimes when value is being derived from this data, is it feeding back to some of the farmers or the, um, the people within the pipeline that are producing some of this data? So value sharing in, in the, in the sense that do they, do, do they get values from the data?
[00:34:23] Ahmad Raji: It could be– It could also be like reduce cost in using the models. If a data generator- You could be– you could, you could have, um, you could have the, um, the opportunity to use the model at a reduced cost. That would also boost the, um, the willingness to even produce more data into, into the system. Great.
[00:34:39] Ahmad Raji: So, um, and then there’s also the ethical considerations, like the government should also come in place to put some policies in place to make sure that data is being used as intended. That’s a good bridge over to Femi. So what, what are the governments doing around this topic of, uh, you know, model and data sovereignty and transparency, accountability, uh, you know, [00:35:00] shared benefits?
[00:35:01] David Bergvinson: Uh, what are you seeing on the policy landscape? There are strategies, but how are these actually getting translated into legislation that, um, preserves that trust and, and ensures value back to the farmer for their data? Yep. Um, just like your last line, I think value for pharma is very critical. But I think that if LLMs need to succeed within the, um, African context, and I mean globally as we’ve seen, um, they need to try to fit within, you know, what I call the laws and frameworks of countries where they’re used.
[00:35:37] Femi Royal: Um, you know, there needs to be very strong compliance and, um, regulation on data protection and alig-aligning this with, um, digital, uh, policies generally. A-and we can see that there’s quite a lot of policy, um, adjustment or would I say framework within, [00:36:00] um, globally around that is governing the use of AI.
[00:36:05] Femi Royal: Um, I still see that Africa or African countries generally lag behind when it comes to building a framework within which AI and LLM should operate. Um, we can see a lot of countries like Canada who has, um, you know, published policies around responsible AI adoption for social impact. Um, there’s a lot of advancement from the OECD as well, where they’ve created like an AI principle.
[00:36:36] Femi Royal: Um, the EU as well has created its AI Act, um, and has done like a bit, a bit of a comprehensive risk-based AI law, uh, within the economy. Um, so I think to a large extent, yeah, to some extent, the African Union is trying its bit to create, um, what I would call its continental AI strategy. Uh, and, and I think the [00:37:00] focus has largely been around, um, ethical, responsible, and equitable AI development across all the member states, you know.
[00:37:08] Femi Royal: Um, and, and this is sort of somewhat in tandem with what the UN as well, the United Nations is doing around creating, um, an AI advisory body, um, just to coordinate collaborations and sort of create some form of standardization and ultimately manage the, um, risk that comes from getting, uh, adopting AI across board.
[00:37:32] Femi Royal: Um, for me, I think there’s much work to do in, in, with- within the African, uh, landscape, um, to not just, um, strengthen public policy, uh, not just strengthening governance framework around the use of AI, but also even spend more eff- uh, put more effort into, um, AI literacy. Um, you know, such that, uh, from a public-private sector [00:38:00] engagement, AI applications are more transparent and accountable and sort of support social good, and there’s a responsible usage of it and not just, um, you know, uh, from an extractive point of view.
[00:38:13] David Bergvinson: Yeah. So, so I– this has been a great conversation, uh, gentlemen. Thank you. Uh, just in closing, what gives you hope or optimism that, uh, large language models can indeed serve the needs of smallholder farmers, especially in Africa? So just in a couple sentences, what, what gives you that optimism that Africa’s on a path to successfully apply LLMs?
[00:38:36] David Bergvinson: So let’s start with you, Femi Yeah. Um, to a large extent, I, I mean, looking through the research that we, we’ve done, um, I think the, the future of AI in, in Africa is it’s not largely about the sophistication of the model or the models that are being created. Um, I think it will be more [00:39:00] about building like a trusted ecosystem.
[00:39:03] Femi Royal: And I think that with the conversations that we’re having, with the papers that we’re putting out there, and with the way we’re, we’re stoking the community to continue to put this at the front burner, um, the advancements in AI capabilities, um, will start to mirror that in a way that models can now become, um, not just powerful, efficient, but it will be more accessible and accessible at pace.
[00:39:32] Femi Royal: Um, and, and you, you would agree with me that a couple of years ago that that’s somewhat unimaginable. So as we look at the realities that sort of impact small scale farmers within Africa, um, and all the challenges that we’ve highlighted in the course of this conversation, uh, we can start to see that, you know, um, some of these conversations are helping to guide government, um, intervention, also guide private sector [00:40:00] engagement to ensuring that, you know, we’re building from a context specific, um, um, perspective.
[00:40:07] Femi Royal: And we’re also eliminating all the bottlenecks and everything that is gonna prevent farmers from actively getting involved with the use of AI to, uh, you know, get benefits that will also take the, um, food security of the sec- of the continents forward. Um, I believe that, uh, there’s so much hope for, and it’s already started.
[00:40:31] Femi Royal: And, and I’m– I believe we, we’re not just gonna be playing catch up as things go. We’re definitely going to be, um, involved actively in, in all of this. Mm-hmm. And, you know, just in closing, uh, Malcolm and, um, just some, some, some points that, uh, also give you optimism, uh, based on your work on this discussion paper of, of models.
[00:40:52] David Bergvinson: So, uh, what’s sort of your parting words for the community and also an invitation for them to contribute to your [00:41:00] discussion paper on this topic as well. So Malcolm Yeah. I think, um, one of the things the ecosystem, particularly the AgTech, um, um, solution providers should leverage on is the trust, right?
[00:41:17] Malcolm Durosaye: It’s the currency for this AI era. And, um, you know, if they can externalize some of those, um, uh, responsibility of actually checking and also doing their own diligence, um, to ensure that there’s, um, a framework that would actually guide the users, uh, you know, it could help, you know, in terms of scaling some of those solutions that they deploy across board.
[00:41:46] Malcolm Durosaye: And also, one of the things we are also advocating for is a multi-stakeholder, um, collaborations where everybody, every actor in the, uh, ecosystem come together just to ensure [00:42:00] that at the end of the day, um, some of these solutions or, you know, all the solutions are actually benefiting, you know, the right people, which are the farmers, and also protecting them against all the extractive, um- Yeah
[00:42:14] Malcolm Durosaye: you know, activities by the external players. So that- Yeah, and that’s a great call-out for the AgX AI community for people to get involved, uh, to ensure their voice is heard and they can contribute to provide a more balanced and complete perspective of the application of AI, so thank you for that. Um, Ahmad, your final words as an author of this discussion paper on models.
[00:42:37] David Bergvinson: What, uh, what gives you hope, uh, based on your research so far? Yeah. Um, thank you. For me, um, what– in my, in my opinion, I think the opportunity is not just in creating a model that will catch up with global AI, it’s more about building agricultural AI system that is grounded in, in local realities. And a good news is that we are [00:43:00] already getting there gradually.
[00:43:00] Ahmad Raji: We’ve seen examples with the likes of Achiko, which is built on, um, the multi-stakeholder, um, ecosystem that Malcolm had mentioned earlier, where they, um, where they consult with farmers, aggregators, input providers, and, and, and the likes. And they’ve created the system to, um, be able to accept input and output in different local languages.
[00:43:20] Ahmad Raji: For instance, in Nigeria, we have the Hausa, Igbo and Yoruba, and then it can even prompt you to even videos in those languages for you to even have a better description of what the AI is trying to explain to you. So we already having tractions. What we– One of the things we need is, um, investment in these, in these systems.
[00:43:36] Ahmad Raji: Then, um, uh, service providers such as internet connectivity, SMS, probably reduce their cost and things like that. So engagement with all of the stakeholders, and everyone has a role to play because we all know that, um, food is actually very important in our lives. So we all have a role to play to make sure that this comes into lives and, um, our efficiency when it comes to farming in Africa and India is, um, is top-notch.
[00:43:59] David Bergvinson: Great. Well, thank you [00:44:00] Femi, Malcolm, Ahmad. Uh, great conversation on models. Uh, it’s also a, a great opportunity for me just to call out that, uh, we invite everyone watching this podcast, if you haven’t subscribed, please do so, so that you can follow future episodes, but also contribute to the development of these discussion papers so that your voice is also heard on these core topics like models.
[00:44:21] David Bergvinson: Uh, you know, we talked already about data in a previous episode. Our next episode, episode four, we’ll talk about benchmarking and how we compare these models against each other to make sure that they’re relevant and offering value to end users. So don’t miss out on that episode for benchmarking and, uh, to do so, just hit the subscribe button below.
[00:44:42] David Bergvinson: Also, we invite you to add comments. What topics are really important for you? Uh, in addition to the ones that we were already covered, there’s many other topics relevant for the responsible use of AI in the agriculture sector. So invite you to let your voice be heard also in the comment section. So please, uh, contribute in that way as well.[00:45:00]
[00:45:00] David Bergvinson: So gentlemen, thank you so much for this conversation on models. Uh, certainly we’re on the start of a very exciting journey on the application of AI to empower farmers, and thank you for each of your roles in making that a reality.
